Bio-medical Ontologies Maintenance and Change Management

Mining Clinical, Immunological, and Genetic Data 221
discretized the age information in buckets: the first bucket ranging from 1 to
30 years, and the following ones having a size of 5 years.
An important problem addressed in this phase has been that of homozygous
alleles, i.e., the situation in which the two alleles for a given locus are
identical. Since we are dealing with information structured in transactions,
i.e., plain sets and not bags, no duplicate items are admitted. Therefore, in
order not to lose information about homozygous alleles, we have created a
special item for each locus indicating the presence of homozygous alleles.
Following the physicians’ indications we have then divided the pathologies
in two main groups, with a total of three item codes:
• Hepatic cirrhosis due to viral infection:
– one item code for HCV+;
– one item code for HBV+;
• Hepatic cirrhosis with non viral origin (mainly autoimmune): one unique
item code for cryptogenetic cirrhosis, primary biliary cirrhosis, necrosis
VBI and VBE, alcoholic cirrhosis, HCC, etc.
Example 1. Table 1 shows a sample of the input file accepted by the algorithm.
The file is composed by several transactions (sets of integers), that can have
different length: a patient can have different types of cirrhosis at the same time
(this is the case of the patient in the third transaction, which has two disease
codes, 1002 and 1004), or it could have some genetic variables missing.
The first transaction in Table 1 regards a female patient, up to 30 years old,
with autoimmune cirrhosis (code 1004) and the following HLA characterization:
A1 =24,A2 =25,B1 =8,B2 =18,DRB11 = DRB12 =1.Here
the DRB1 is a homozygous locus, i.e. it has the same value in each of the two
chromosomes; this is represented by the presence of the item code 2499.
We ran the software on the input database and produced the frequent closed
itemsets. After the mining phase we performed a post-processing phase, in
which the extracted patterns were automatically selected w.r.t. their interestingness.
Here by interesting pattern we mean a pattern with an exceptionally
high frequency in the patients database w.r.t. the frequency of the
same pattern, without disease information, in the control database. In fact,
this situation would correspond to a possible association between a specific
pattern and a specific class of diseases.
Table 2 reports some of the results obtained with this method. In the
second column we have the pattern which if taken together with the disease
code in the third column, constitutes a frequent closed itemset in the patients
Table 1. A sample of the input database
2 30 2024 2025 2208 2218 2401 2499 1004
1 30 2001 2002 2203 2297 2403 2499 1004
2 35 2024 2031 2214 2251 2414 2404 1004 1002